Re-imagining Vivaldi with HTML5/WebGL. PhD Candidate: Djuro Mirkovic Supervisors: Prof. Grenville Armitage and Dr. Philip Branch

Transcription

1 Re-imagining Vivaldi with HTML5/WebGL PhD Candidate: Djuro Mirkovic Supervisors: Prof. Grenville Armitage and Dr. Philip Branch Centre for Advanced Internet Architectures (CAIA) Swinburne University of Technology Outline Problem Scenario Motivation Goal Network Coordinate System (NCS) Model Space Vivaldi Discussion and Examples Demo Summary 15 October

2 Problem Scenario Client wants to select a Server with the lowest round trip time (RTT) Measure RTT with all Servers? Wastes resources (eg. Bandwidth) May overflow NAT Table poor performance May take a long time to gather all RTT measurements After time t, RTT measurements are stale dmirkovic@swin.edu.au 15 October Motivation Peer-to-peer (P2P) applications treat the Internet as a black box Client-Server model optimal selection can be improved Scenarios: Network Coordinate System (NCS) with large number of nodes or small data size can improve performance and resource allocation dmirkovic@swin.edu.au 15 October

3 Goal A model that is able to predict the distance between two Nodes, given only a few measurements. dmirkovic@swin.edu.au 15 October Network Coordinate Systems (NCS) Centralised Predict using Synthetic Coordinates as described in Global Network Positioning (GNP) [1] Requires Landmark(s) (most likely being Server Nodes or accurate Nodes) Decentralised Predict using Network Coordinate System as described in Vivaldi [2] Landmark(s) are optional [1] - [2] dmirkovic@swin.edu.au 15 October

4 Model Space N-dimension Euclidean Model At least 2D Model is required Increasing N-dimensions substantially in Euclidean Model does not have significant improvements May have an opposite undesired effect increasing the computational power required for higher N-dimensions Using a 3D to 5D Model is adequate 2D + Height Vector Model Spherical Model Cylindrical Model dmirkovic@swin.edu.au 15 October Decentralised - Vivaldi Vivaldi is solving a simultaneous equation distributed in time and space Nodes iteratively update their own coordinates based on current network state A Local Node that is communicating with Remote Node(s) requires a very small sample of Remote Node(s) to get an overall state Nodes communicating can piggyback on existing application traffic Adds very little overhead dmirkovic@swin.edu.au 15 October

5 Springs Overestimate and Underestimate RTT Difference between measured and calculated RTT Springs compress and stretch to move Node D to new coordinate space after n iterations of Vivaldi Push and Pull effect (Hooke s Law) dmirkovic@swin.edu.au 15 October Triangle Inequality NCS assumption is that Triangle Inequality holds for Node paths. The sum of the lengths of any two sides, must be greater than or equal to the length of remaining side. [1] Eg. bc ab + ac [1] - dmirkovic@swin.edu.au 15 October

6 Vivaldi Example 1 (Step 1) Local Node A will do sequence of communication with other Remote Nodes (B and C) dmirkovic@swin.edu.au 15 October Vivaldi Example 1 (Step 2) A receives B s Coordinates (-20, 15) and RTT (30ms) A calculates distance to B (50ms) RTT A, B = (A x B x ) 2 +(A y B y ) 2 RTT A, B = (A x B x ) 2 +(A y B y ) 2 dmirkovic@swin.edu.au 15 October

7 Vivaldi Example 1 (Step 3) Overestimated: A needs to be pulled towards B dmirkovic@swin.edu.au 15 October Vivaldi Example 1 (Step 4) A (over n-iterations not instantly!) approaches B s measured RTT. A adjusts coordinates (A x, A y ) dmirkovic@swin.edu.au 15 October

8 Vivaldi Example 1 (Step 5) A communicates with C Note: coordinate position of A has been moved (when A completed one iteration of Vivaldi with B) dmirkovic@swin.edu.au 15 October Vivaldi Example 1 (Step 6) A receives C s Coordinates (50, -25) and RTT (40ms) A calculates distance to C (20ms) RTT A, B = (A x B x ) 2 +(A y B y ) 2 dmirkovic@swin.edu.au 15 October

9 Vivaldi Example 1 (Step 7) Underestimated: A needs to be pushed away from C dmirkovic@swin.edu.au 15 October Vivaldi Example 1 (Step 8) A (over n-iterations) approaches C s measured RTT. A adjusts coordinates (A x, A y ) dmirkovic@swin.edu.au 15 October

10 Vivaldi Example 1 (Step 9) Previous slide, the distance between A and B now would be wrong Vivaldi process will iterate until all Local and Remote Nodes measured RTT have same calculated distance (or RTT) dmirkovic@swin.edu.au 15 October Vivaldi Landmarks Example 2 (Step 1) 13 Stable Landmark Nodes (A to M) calculated their coordinates (over n-iterations). New Node N joins the process dmirkovic@swin.edu.au 15 October

11 Vivaldi Landmarks Example 2 (Step 2) 2D Euclidean Model requires at least 3 Remote Node samples, otherwise many solutions may exist in NCS Number of Remote Node samples is equal to number of dimensions + 1 (3D would need 4, 4D would need 5, etc) N needs 3 Remote Node samples: Does N select 3 random? Does N select more than 3 random? If Landmarks used: Does N select 3 shortest RTT measured? Does N select 3 longest RTT measured? Does N select 3 different geographical locations? dmirkovic@swin.edu.au 15 October Vivaldi Landmarks Example 2 (Step 3) N randomly selected Remote Nodes (E, F, and M). Measured RTT from E, F, and M are 200ms, 150ms, and 30ms respectively dmirkovic@swin.edu.au 15 October

12 Vivaldi Landmarks Example 2 (Step 4) N after n-iterations of Vivaldi determines coordinates and satisfies the computed RTT to be the same as measured RTT dmirkovic@swin.edu.au 15 October Vivaldi Landmarks Example 2 (Step 5) RTT between N and Remote Nodes (A, B, C, D, G, H, I, J, K, and L) can be predicted by the Euclidean Distance between their coordinates without direct measurement dmirkovic@swin.edu.au 15 October

13 Vivaldi Landmarks Example 2 (Step 6) N best option is C (25ms) N calculates the predicted RTT for A, B, C, D, G, H, I, J, K, and L faster than it takes to measure the RTT for each Remote Node Previous scenario had only 13 Remote Nodes, what if we had a large number of Remote Nodes (10,000, 100,000, 1 million, etc)? dmirkovic@swin.edu.au 15 October Vivaldi space rotation Each Node in the Vivaldi process starts at the origin, and obtains a random starting position Vivaldi with different seeds (varying random starting position) using the same experimental data, an interesting observation is that: the overall stable position in NCS in one seed universe becomes nearly like a rotation in the second seed universe dmirkovic@swin.edu.au 15 October

14 Vivaldi Caveats Triangle Inequality violations Malicious Nodes can affect the calculated coordinates 15 October Demo of 3D Vivaldi Simulations Hard to conceptualise the overall Vivaldi process without a suitable simulation tool dmirkovic@swin.edu.au 15 October

15 Summary Vivaldi is a decentralised NCS 3D 5D Euclidean Model is an acceptable model to use in NCS Vivaldi can optimise application performance and resource allocation dmirkovic@swin.edu.au 15 October Acknowledgement Thank you very much to Prof. Grenville Armitage and Dr. Philip Branch for their supervision support throughout my PhD. dmirkovic@swin.edu.au 15 October

16 Questions? 15 October